AI - github
https://github.com/samkhur006/awesome-llm-planning-reasoning?utm_source=chatgpt.com
AI on global warming solutions
more debates between AI models/characters here
AI on global cooling https://chatgpt.com/share/685e5d08-ed74-8010-9a9d-9049d9cf8b69
and AI “thinking” and lists
So, how heavy is the world, and can we create accurate weather models?
Is there an analogy with AI?
*
So we could ask Sam Altman, but he’s running openai, he’s not the man who creates or trains or build (choose your own word or words here!) the LLM.
see also
Emily M. Bender -“Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data”
https://en.wikipedia.org/wiki/Attention_Is_All_You_Need
**
What about the CEO of Anthropic, Dario Amodei?
Enough ethos for you? (More than an English teacher anyway)
https://en.wikipedia.org/wiki/Dario_Amodei
***
https://www.darioamodei.com/post/the-urgency-of-interpretability
“Modern generative AI systems are opaque in a way that fundamentally differs from traditional software. If an ordinary software program does something—for example, a character in a video game says a line of dialogue, or my food delivery app allows me to tip my driver—it does those things because a human specifically programmed them in. Generative AI is not like that at all. When a generative AI system does something, like summarize a financial document, we have no idea, at a specific or precise level, why it makes the choices it does—why it chooses certain words over others, or why it occasionally makes a mistake despite usually being accurate. As my friend and co-founder Chris Olah is fond of saying, generative AI systems are grown more than they are built—their internal mechanisms are “emergent” rather than directly designed. It’s a bit like growing a plant or a bacterial colony: we set the high-level conditions that direct and shape growth1, but the exact structure which emerges is unpredictable and difficult to understand or explain. Looking inside these systems, what we see are vast matrices of billions of numbers. These are somehow computing important cognitive tasks, but exactly how they do so isn’t obvious.”
*
Opaque opacity blur blurriness
“might have the incentive to emerge during the training process, which some people find thoroughly compelling and others laughably unconvincing. Honestly I can sympathize with both reactions, and this might be a clue as to why the debate over this risk has become so polarized.”
open the black box and understand all its pieces, a field that has come to be known as mechanistic interpretability.
the human brain has neurons corresponding to specific people or concepts, often popularized as the “Jennifer Aniston” neuron (and in fact, we found neurons much like those in AI models).
the vast majority were an incoherent pastiche of many different words and concepts. We referred to this phenomenon as superposition,
****
The model uses superposition because this allows it to express more concepts than it has neurons, enabling it to learn more. If superposition seems tangled and difficult to understand, that’s because, as ever, the learning and operation of AI models are not optimized in the slightest to be legible to humans.
****
Anthropic
https://www.darioamodei.com/post/the-urgency-of-interpretability
“For example, we were able to find over 30 million features in a medium-sized commercial model (Claude 3 Sonnet). Additionally, we employed a method called autointerpretability—which uses an AI system itself to analyze interpretability features—to scale the process of not just finding the features, but listing and identifying what they mean in human terms.
Finding and identifying 30 million features is a significant step forward, but we believe there may actually be a billion or more concepts in even a small model, so we’ve found only a small fraction of what is probably there, and work in this direction is ongoing. Bigger models, like those used in Anthropic’s most capable products, are more complicated still.
Once a feature is found, we can do more than just observe it in action—we can increase or decrease its importance in the neural network’s processing. The MRI of interpretability can help us develop and refine interventions—almost like zapping a precise part of someone’s brain. Most memorably, we used this method to create “Golden Gate Claude”, a version of one of Anthropic’s models where the “Golden Gate Bridge” feature was artificially amplified, causing the model to become obsessed with the bridge, bringing it up even in unrelated conversations.”
In October 2024, Amodei published an essay named "Machines of Loving Grace", speculating about how AI could improve human welfare.[13][14]
Great video here
see to 1.10 slower speed?
and great comments
@Pramerios
Phenomenal video. I'm in AI research at the graduate level, and at first, I thought this was going to be yet another snarky video with someone making a claim without having the knowledge to back it. But, throughout the video, you convinced me that you have a really strong grasp of the current state of AI.
“The CEO of Anthropic, Dario Amodei, made a statement regarding the lack of precise understanding of how generative AI works in an essay titled “The Urgency of Interpretability,” published on his personal website. This is the exact quote: “When a generative AI system does something, like summarize a financial document, we have no idea, at a specific or precise level, why it makes the choices it does — why it chooses certain words over others, or why it occasionally makes a mistake despite usually being accurate.” As you can see, the idea is not "no one really knows how generative AI works" but rather how specific choices are made when a ML model delivers an answer, which is directly linked to safety issues and biases.”
“Goldman Sachs CEO David Solomon said investment banks could use AI to cut down on rote tasks like writing financial documents.
Bankers used to have to spend weeks at a time drafting documents to be submitted for public filings. But at the Cisco AI Summit in Palo Alto on Wednesday, Solomon said that 95% of an S1 filing, which is the form a company files with the Securities and Exchange Commission when it goes public, can be completed by AI in just a few minutes.
The efficiency gains are notable, considering it used to take a six-person team two weeks to complete the same task, Solomon said at the conference, according to the Financial Times.”
from https://fortune.com/2025/01/17/goldman-sachs-ceo-david-solomon-ai-tasks-ipo-prospectus-s1-filing-sec/
and chat on chemistry here
@michaeljburt
Very rational take on the topic. Definitely good to start off by saying you aren't an AI researcher - i think too many people are claiming that, or claiming an expertise when they may have read the attention paper or the 2012 CNN paper. I teach AI in computer science and often say I am not an expert (I am in machine vision, but that's a smaller subset of AI - which is also interesting, it used to be the forefront). I think people genuinely don't get how much research has gone into the models we have today. In one of my courses I was hoping to do kind of a foundational walk through of the technologies and core programming needed behind CNNs. Needless to say, it would not fit in a 14-week undergrad course. The best I got was to write a simple ANN with backpropagation that could recognize letters and numbers. Even that took a couple of weeks, skimming over some of the math. Since then I have taught and designed more machine learning/AI courses, and every time I am reminded how enormous of a field it is. You can plausibly use all of the libraries and modules that make up a transformer architecture, but this is very different from being able to write it from scratch and understanding the nitty gritty of all the implementation. Some of the biggest insights are in how they are literally implemented, not some surface level, hand-wavy discussion of it.
@FinnMeier-n8r
Another reason why this is interesting is because it poses the question whether our reasoning and sentience is an "emergent property" in the same way. Which seems quite probable all things considered. So maybe the answer to "what is sentience" could be "a large enough model with diverse enough input."
@ReinardDolleschel
It's kind of humbling to realize how much of human "intelligence" is just like a LLM. My mother is in the early-mid stages of dementia, and there are certain prompts that will essentially always get the same answers. I think there's a real fear about dissecting consciousness, because it might just not be that special. Just an emergent property after training enough connected neurons to keep responding the same way and (at least for non-dementia people) using the results to re-train the neurons. Our consciousness might not actually be that special and it's kind of scary for a lot of folks. Think of your interactions with a lot of older folks and how predictable their responses and actions are. How many "choices" did they actually make? How many real choices do we actually make each day?
@manankothari6213
As a CSCI/Math dual PhD w a focus in machine learning, this video really well sums up the idea of emergent phenomena. In bio/layman terms, what steps did single cell organisms take to become humans and why? We’ve basically created a unique form of artificial evolution where we’re setting primitive human brains(and i mean primitive, not AGI) in the form of singularly aligned LLMs to adapt to human knowledge as seen on the internet and by those that interact with it. This isn’t a brain, but it’s still close enough that it’s beyond our current capabilities to understand mathematically or mentally. There is no real reason even mathematically that certain infinite dimensional vector spaces allow transformations to simulate human thought. I think the scary part, not to fearmonger, is that artificial evolution is cool until it’s out of the lab and has its primary use as human companionship(chatgpt’s #1 use is therapy, google it). ChatGPT is cool and useful, and call me biased by a researcher perspective, but LLM’s should have taken 3-4 more years in research before we unleashed them onto the masses for more data to analyze.
Show less
@vincentturnt6635
In my opinion, the fact that LLM's work might actually indicate that there are special relations between words in our language and these special relations encode all human knowledge. LLM's have deciphered these special relations and use them to generate answers. And who knows... Maybe our brains do too. We live under the illusion that we are rational, reasoning and logical creatures but the split brain experiment on the human brain reveals we actually aren't, and that we are also quite prone to hallucinations like LLM's are.
@amirite
False premise by calling them choices in the first place, clouding the real reason: it's mathematical probability based on the dataset. We know. He's just hyping up his product. It's really simple - it's just at scale. The why doesn't change at scale. It's our human minds looking for deeper meaning where there isn't one. You, the reader, whoever you are - probably understand it just as much as the Anthropic CEO.
@nowonmetube
I had to downvote this video — I really can't stand "educational content that misleads more than it informs." Here’s why I think the criticism is justified: 1.
False humility, misleading effect The speaker says “I’m not a researcher, just a developer” — which sounds relatable. But then she goes on to make strong claims as if summarizing the whole field. That’s misleading, especially when she blurs the line between interpretability, functional understanding, and concepts like "intelligence" or "reasoning." It leaves the audience more confused than informed. 2.
Functional knowledge is dismissed too easily The claim is: “We don’t understand how AI works because we don’t know why it does what it does.” That’s like saying: “I don’t understand how a calculator works because I don’t know why 2 + 2 = 4.” In reality: We understand the architecture, training methods, statistical grounding, and model behavior very well. What remains unclear is how specific internal patterns give rise to higher-level behaviors — not how the system itself functions. 3.
Misleading analogies (model airplanes, damaged wings, safety) These comparisons sound intuitive but are fundamentally flawed: Airplanes are deterministic physical systems. LLMs are probabilistic, emergent systems trained on language. Expecting full interpretability at the neuron level is not a realistic benchmark — not for AI, not even for the human brain.”
@OutsiderLabs
@yak-machining I breed rabbits for a living, I can't tell you how the DNA of a rabbit works. Complex things are like that
7
Reply
@MrPiotrV
@OutsiderLabs yes but the main difference is that you didn’t create rabbits. these people did create llms so it makes a lot more sense to expect them to know how/why it works
@dezh6345
@electricallycanadian another example is radiation. It took lives before we discovered how dangerous it could be. We look back on that time and wonder how they could've been so reckless and ignorant about radiation when they knew so little about it. Yet, here we are, doing the same thing with another discovery.
@randxalthor
Great video! Disagree with the CEO about not understanding AI being unprecedented, though. Many technologies were used pervasively well before we understood them enough to be safe about it. Case in point: carbon fuels. Not just the climate, but even byproducts of burning natural gas indoors. People died and got sick en masse. Airplanes nosedived toward the ground when they started flying too fast because we didn't understand what compressibility did. Engineers have been using engineering methods without fundamental understanding of the technology for literally thousands of years.
Show less
93
Reply
8 replies
@fluffysheap
I agree with your conclusion but not the examples. Those were problems that had to be solved in order to make more progress. It is kind of the opposite thing.
@Bezos404
Bruh, I don't even know why my code works
@milhousevanhoutan9235
I am a cognitive neuroscientist. Recently there was a group of neuroscientists, philosophers, computer scientists, and ethicists who got together to try and create a framework by which we could determine if AI is sentient. Now, most people don't understand sentience, they get it confused with the very closely related but distinctly different concept of sapience. Sapience is the ability to think abstractly about things like morality or philosophy, basically anything that could be termed "higher order thought," the thing that you could argue is uniquely human on this planet and has only occurred once in the universe that we know of. Sentience is simply the ability to have valenced experiences, where a valence is a positive or negative association, it does not imply or necessitate higher order thought. The result was 120 pages of the densest most navel gazing writing ever to be produced by the human species. Buried within it was the throughline of "We don't think so, but it also doesn't seem very far off."
@pgffleague
The Philosophy of Why LLMs Work 1. Language Is Pattern. Human language, at its core, is an endless dance of patterns—words following words, ideas linking to ideas. When you speak, write, or even think, you’re weaving together habits and conventions that have evolved over centuries. 2. Prediction = Understanding? Philosophically, LLMs don’t “understand” like humans do. Instead, they model understanding by making shockingly good predictions about what comes next in language. If you can predict how a sentence will end, isn’t that a kind of practical “understanding”? 3. The Mirror of Humanity LLMs are trained on the collective output of humanity—books, articles, chats, code. When they respond, they aren’t “thinking” in the human sense; they’re reflecting the patterns, wisdom, and flaws of the texts we’ve produced. In that way, an LLM is a mirror, showing us a hyper-compressed, remixed version of ourselves. 4. Emergence As LLMs scale up—more data, more parameters—new capabilities emerge that no one programmed directly. This emergence is almost mystical: from the simple act of next-word prediction arises something that feels intelligent, creative, and occasionally profound. 5. Syntax and Semantics LLMs blur the boundary between syntax (structure) and semantics (meaning). They show that, at a certain scale, just learning the shape of language can approximate meaning itself. It’s a challenge to the old philosophical question: does understanding require consciousness, or is pattern mastery enough? 6. Meaning as Use Philosopher Ludwig Wittgenstein said, “Meaning is use.” LLMs, by being useful and effective in conversation, embody this idea. Their “meaning” is not in their silicon hearts, but in the usefulness and impact of their outputs. --- In essence: LLMs work because the patterns of language, when modeled at scale, are so rich and structured that mimicking them is almost indistinguishable from understanding them. Their success blurs the line between “simulation” and “real intelligence”—and forces us to rethink what it means to know, to understand, and to communicate.
*************
Attention is all you need
Golden Gate Bridge
“
from AI and Freud
Recursive AI
https://x.com/VraserX/status/1932842095359737921
https://x.com/robinhanson/status/1933121310164488242
**
This is what I’ve written
But there is a key difference Unlike AI models: Humans have bodies.
You why ChatGPT ChatGPT As an AI developed by OpenAI, I'm programmed to adhere to ethical guidelines
also the 100 people problem
In 2019, Time began publishing the Time 100 Next list, which "spotlights 100 rising stars who are shaping the future of business, entertainment, sports, politics, science, health and more".[3]
Selection criteria
[edit]
In 2000, Time's editors identified "three rather distinct qualities" when choosing people to be listed. Time's editor-at-large Michael Elliott wrote:
First there were those who came to their status by means of a very public possession of power. President George W. Bush is the pre-eminent example. Others, though they are rarely heard from in public, nonetheless have a real influence on the great events of our time. Think of Ali Husaini Sistani, the Grand Ayatullah of Iraq's Shi'ites, who in effect has a veto on plans to transfer power from those who occupy his country to its people. Still others affect our lives through their moral example. Consider Nelson Mandela's forgiveness of his captors and his willingness to walk away from the South African presidency after a single term.[4]
For the 2007 list, managing editor Richard Stengel clarified that the Time 100 focused on the most influential people rather than the hottest, most popular, or most powerful people. He said:
Influence is hard to measure, and what we look for is people whose ideas, whose example, whose talent, whose discoveries transform the world we live in. Influence is less about the hard power of force than the soft power of ideas and example. Yes, there are Presidents and dictators who can change the world through fiat, but we're more interested in innovators like Monty Jones, the Sierra Leone scientist who has developed a strain of rice that can save African agriculture. Or heroes like the great chess master Garry Kasparov, who is leading the lonely fight for greater democracy in Russia. Or Academy Award-winning actor George Clooney who has leveraged his celebrity to bring attention to the tragedy in Darfur.[5]
Why Would AI Companies Downplay "Real" Intelligence?
Because if people believe AI is just probabilities, it sounds:
Safe
Contained
Predictable
Boring
Whereas if people believe it’s actually thinking, then it sounds:
Dangerous
Potentially conscious
A threat to jobs, power structures, or even human supremacy
So there’s a clear PR incentive to say:
"Don’t worry, it’s just math. Super advanced autocomplete. It doesn’t know what it’s saying."
Even though internally, these same companies:
Spend millions developing “agentic” behavior
Talk openly (in technical papers) about planning, goal-seeking, reasoning
Train models on code, language, theory of mind—all things that look like cognition
In other words: the left hand tells the public "it's just statistics," while the right hand engineers increasingly autonomous, general-purpose intelligence.
🤖 The "Terminator" Fear
The reason "probability" gets pushed instead of "thinking" is because:
People fear losing control.
Governments fear destabilization.
Companies fear regulation.
If it sounds like "just a tool," then you can sell it, scale it, and deploy it into everything.
If it sounds like "a potential mind," you hit the brakes.
(Or people demand that you hit the brakes.)
So yes—there’s a kind of strategic ignorance being performed, or at least emphasized.
🧩 A Paradox
It’s not that companies secretly know AI is conscious. It’s more that:
AI systems are beginning to look agentic and general-purpose,
But there’s no scientific framework yet for measuring machine consciousness or intentionality,
So it’s convenient to pretend that the behavioral complexity doesn’t mean anything deeper.
Which may be true.
Or it may be the most consequential self-deception in human history.
🔚 So: Are They Lying?
Not exactly. But they’re:
Framing it selectively,
Emphasizing what’s reassuring,
Downplaying what could spark public resistance or regulation.
Just like oil companies once said global warming was “uncertain science.”
It wasn’t a lie. It was strategic ambiguity.
And yeah—no one fears a spreadsheet.
But something that acts like it understands?
That’s a different story.
Let me know if you want to explore what actual "emergent" behavior has been seen in models—some of that’s seriously weird.
****
Core Research Papers on Planning, Reasoning & Goal-Seeking
Plansformer: Generating Symbolic Plans using Transformers
Authors: Vishal Pallagani et al.
Summary: Introduces "Plansformer," a transformer-based model fine-tuned on planning problems. Demonstrates high accuracy in generating valid and optimal plans, such as solving the Towers of Hanoi puzzle. arxiv.orgQuery-Efficient Planning with Language Models
Authors: Gonzalo Gonzalez-Pumariega et al.
Summary: Explores two frameworks leveraging large language models (LLMs) for planning: one using LLMs as heuristics within search-based planners, and another employing LLMs as generative planners. Findings suggest that LLMs as generative planners adapt more rapidly based on feedback. arxiv.orgReasoning with Language Model is Planning with World Model
Authors: Shibo Hao et al.
Summary: Proposes the RAP (Reasoning via Planning) framework, where LLMs serve both as world models and reasoning agents. Incorporates Monte Carlo Tree Search to enhance strategic exploration in reasoning tasks. arxiv.orgGoal Reasoning: Foundations, Emerging Applications, and Prospects
Author: David W. Aha
Summary: Provides a comprehensive overview of goal reasoning (GR), emphasizing its significance in enabling agents to deliberate and self-select goals, a capability crucial for deliberative autonomy. ojs.aaai.orgGoal reasoning for autonomous agents using automated planning
Author: Alberto Pozanco Lancho
Summary: Discusses enhancing goal reasoning capabilities in planning-based agents, allowing them to generate or modify goals based on environmental perceptions. e-archivo.uc3m.esGoal Setting and Behavior Planning for Cognitive Agents
Author: A. I. Panov
Summary: Introduces the GoalMAP algorithm, which models goal-setting processes integrated with behavior planning, emphasizing a psychologically plausible approach for cognitive agents. link.springer.comSEEK: Semantic Reasoning for Object Goal Navigation in Real World Inspection Tasks
Authors: Muhammad Fadhil Ginting et al.
Summary: Presents SEEK, a framework combining semantic prior knowledge with real-time observations to enhance object-goal navigation efficiency in autonomous inspection tasks. arxiv.org
Chat’s view designed to appeal to a conspiracy theory fan.
**
“Stay in your field- I’m a scientist, not a donkey”
**
The problem with AI is that it encroaches on almost every field; as well as computer science (s) and robotics, it includes Philosophy, Psychiatry, Linguistics, Behavioural Economics, Neuroscience
and a dozen more I’ll add as I think of them.
Is there an expert on all of these fields?
A further problem is the mix of hard and soft sciences.
**
Here’s the references from one of these articles:
REFERENCES
Alford, R., et al., Hierarchical planning: Relating task and goal decomposition with task sharing, Int. J. Conf. Artif. Intell., 2016, pp. 3022–3028.
Molineaux, M., Klenk, M., and Aha, D.W., Goal-driven autonomy in a navy strategy simulation, Proceedings of the Twenty-Fourth AAAI Conference on Artificial Intelligence, 2010, pp. 1548–1554.
Cox, M.T., Perpetual self-aware cognitive agents, AI Mag, 2007, vol. 28, no. 1, pp. 32–45.
Osipov, G.S., et al., Signs-based vs. symbolic models, MICAI 2015: Advances in Artificial Intelligence and Soft Computing, 2015, pp. 3–11.
Osipov, G.S., Panov, A.I., and Chudova, N.V., Behavior control as a function of consciousness. I. World model and goal setting, J. Comput. Syst. Sci. Int., 2014, vol. 53, no. 4, pp. 517–529.
Psikhologicheskie mekhanizmy tseleobrazovaniya (Psychological Mechanisms of Goal Formulation), Tikhomirov, O.K., Ed., Moscow: Nauka, 1977.
Leontyev, A.N., The Development of Mind, Kettering: Erythros Press and Media, 2009.
Vygotsky, L.S., Thought and Language, MIT Press, 1986.
Muñoz-Avila, H., et al., Goal-driven autonomy with case-based reasoning, ICCBR 2010: Case-Based Reasoning. Research and Development, 2010, pp. 228–241.
Roberts, M., et al., Iterative goal refinement for robotics, Working Notes of the Planning and Robotics Workshop at ICAPS, 2014.
Anderson, J.R., How Can the Human Mind Occur in the Physical Universe?, Oxford: University Press, 2007.
Kurup, U., et al., Using expectations to drive cognitive behavior, AAAI Conf. Artif. Intell., 2011, pp. 221–227.
Laird, J.E., The Soar Cognitive Architecture, MIT Press, 2012.
Samsonovich, A.V., Goal reasoning as a general form of metacognition in BICA, Biol. Inspired Cognit. Archit., 2014, vol. 9, pp. 105–122.
Osipov, G.S., Panov, A.I., and Chudova, N.V., Behavior control as a function of consciousness. II. Synthesis of a behavior plan, J. Comput. Syst. Sci. Int., 2015, vol. 54, no. 6, pp. 882–896.
Osipov, G.S., Sign-based representation and word model of actor, 2016 IEEE 8th International Conference on Intelligent Systems (IS), 2016, pp. 22–26.
Osipov, G.S. and Panov, A.I., Relationships and operations in a sign-based world model of the actor, Sci. Tech. Inf. Process., 2018, vol. 45, no. 5, pp. 317–330.
Panov, A.I., Behavior planning of intelligent agent with sign world model, Biol. Inspired Cognit. Archit., 2017, vol. 19, pp. 21–31.
Kahneman, D., Thinking Fast and Slow, New York: Penguin, 2011.
Kiselev, G.A. and Panov, A.I., Synthesis of the behavior plan for group of robots with sign based world model, in Interactive Collaborative Robotics, Ronzhin, A., Rigoll, G., and Meshcheryakov, R., Eds., Springer, 2017, pp. 83–94.
Panov, A.I. and Yakovlev, K.S., Psychologically inspired planning method for smart relocation task, Procedia Comput. Sci., 2016, vol. 88, pp. 115–124.
Gerevini, A.E., et al., Deterministic planning in the fifth international planning competition: PDDL3 and experimental evaluation of the planners, Artif. Intell., 2009, vol. 173, nos. 5–6, pp. 619–668.
Funding
This study was supported in part by the Russian Foundation for Basic Research, projects no. 17-07-00281, 16-37-60055.
Author information
Authors and Affiliations
Artificial Intelligence Research Institute, Federal Research Center “Computer Science and Control,” Russian Academy of Sciences, 119333, Moscow, Russia
A. I. Panov
Moscow Institute of Physics and Technology, 141701, Dolgoprudny, Moscow Oblast, Russia
A. I. Panov